Thanks. Haven't done a controlled accuracy comparison yet, just tracking token usage and session time. The main win I see is the agent skipping the exploration phase entirely and going straight to the right files.
Whether that improves accuracy or just speed is hard to separate honestly.
The MCP integration has been the most useful part for me personally. get_overview once at the start of a session and the agent already knows the module layout. Beats watching it read 12 files.
If you try it on a client codebase I would be curious to hear how it holds up. Most of my testing has been on open source repos so far.
https://github.com/glincker/stacklit
Archit Mittal
I Automate Chaos — AI workflows, n8n, Claude, and open-source automation for businesses. Turning repetitive work into one-click systems.
This is solving a real pain point. I run AI agents on client codebases regularly and the orientation cost is brutal — agents burning through tokens just to understand project structure before doing any actual work. The 4k token index vs 500k raw context is a massive efficiency gain. The MCP server integration is particularly interesting since that's becoming the standard for tool-agent communication. Have you benchmarked how much this improves first-task accuracy compared to agents that do the full file-by-file scan?